Anthropic on September 18, 2026 published the first edition of its R&D Automation Index, an internal measurement designed to disclose how much of its own model-development work is now being driven by Claude. The headline number: Claude sits at the "leads" tier for 26% of Anthropic's AI R&D, up from under 1% as recently as February 2026.
An AL0-AL5 Ladder For AI-In-The-Loop
The index uses a six-step scale, from AL0 (no AI involvement) through AL5 (fully autonomous). Claude's headline share sits at AL4 – a level Anthropic defines as "the AI completes most tasks end-to-end from a high-level prompt, with human supervision." Anthropic is explicit that the top tier, AL5, is not yet reached: "Claude is not operating fully autonomously for any measured subset of AI R&D work."
30,000 Agents, 1 In 47,000 Blocked
Anthropic also disclosed the raw operational metrics behind the number. Its primary internal platform ran ~30,000 Claude agents simultaneously during August 2026, and its safety-monitoring system blocked roughly 1 in 47,000 agent decisions out of more than a billion total. Roughly 6% of AI R&D compute for the week of July 13–20, 2026 went to safety-focused work; that share rises to 12% when restricted to AI-driven R&D. Anthropic has committed to publishing the index regularly, with third-party verification.
Methodology In Brief
To score the index, Anthropic sampled 20% of staff across its model-R&D departments during July 2026. Claude agents then reviewed the sampled work and broke it into ~15,000 granular tasks arranged across 542 nodes, with human raters checking a subset. Model ratings and employee ratings agreed exactly 59% of the time. That agreement rate is Anthropic's answer to the obvious objection: an AI grading its own R&D deserves a control group.
Why It Matters
The disclosure is the most quantitatively specific look any frontier lab has given at how far AI-in-the-loop model development has actually progressed inside a top-tier lab. It also arrives at a moment when the same firms are asking regulators for wider latitude: it dovetails with the joint FINRA-style safety body Anthropic, OpenAI and Google DeepMind are pushing, and with Anthropic's own Bay Area wet-lab announcement earlier this week. Together those moves paint a picture of a lab arguing publicly that it is still safely in control while quietly pointing out that AI is doing more than a quarter of its own homework.
Reporting based on coverage from Implicator.ai, Enterprise DNA, Technology.org and Anthropic's own R&D Automation Index disclosure.